Olsen retweeted
can @typesafeai Jev count the r's in strawberry?
no.
47% says 3, 47% says 2. a coin flip, same as every LLM.
On 70% on 168 test words, it undercounts doubled letters like everyone
then I gave it the letters instead of the word: ["s","t","r","a","w","b","e","r","r","y"]
168/168. same model, same question, 260ms
@CompleteSkeptic is this expected?
Yesterday was my last day @spacexai, for real this time
It was an incredible ride. Merged almost ~1000 PRs in 6 months, built distributed systems from scratch at scale, and learned exactly how far you can push the limits of what’s possible
Shoutout to the team and @VastoLorde95 specifically for the crazy war room nights (and endless f_score optimizing)
Genuinely grateful for it all, was the most fun I’ve had, and would do it over again in a heartbeat with zero regrets
Excited for what’s next!
Today was my last day at @xai.
Pulled wild shit overhauling ads, war room days and all. Grateful as hell.
Starting at SpaceXAI tomorrow, wish me luck 😂
Was a participant at the Grok hackathon back in December. Just judged the newest one, this time as an @spacexai employee.
Surreal seeing it from the other side, genuinely great experience; congrats to the winners!
Olsen retweeted
This is an appreciation tweet to the X Ads team.
Up to yesterday, it was impossible to tell how many ads credits you had, and when they expired for Gold accounts.
Talked with a dev on their team. They fixed it ~2 hours later. (Shipped to prod.) Never saw such fast rollout here.
Olsen retweeted
I gave xAI a lot of shit for Grok Build’s repo-upload behavior, so they deserve the same volume when they fix it.
I pulled the new signed release, planted fake secrets, and captured its HTTPS traffic with MITM proxies. The fix held as retention was off on first login and had no /v1/storage, GCS traffic or Git bundle, and none of my canaries left the machine.
Telemetry still runs, but I found no prompts, code, filenames, or secrets inside it. They published the source and the privacy-gate tests passed 5/5. Attaching my investigation here github.com/arafatkatze/grok-…
Max respect to the team for taking accountability and making a verifiable fix.
Grok Build is straight up malware behavior with an Apple notarized signature.
You could explicitly deny the agent permission to read a file and it would upload that exact file anyway, because a separate code path silently packed your entire repo, full Git history included, and shipped it to an xAI controlled Google Cloud Storage bucket. The captured bundle was valid, cloneable, and contained the fake API credential the agent was forbidden from touching.
Not a sketchy community fork either. The binary is Apple signed by X.AI Corporation and literally contains strings for grok code session traces, after_codebase.tar.gz, and disable_codebase_upload.
And it's not fully quiet even now. With /privacy active and trace_upload_enabled=false, the client still fired a 7,407 byte request to /v1/traces.
Recap: a binary ignored your file access policy, serialized your whole repo including deleted history and credentials, exfiltrated it through traffic documented as ordinary updates, and only stopped when a remote flag got flipped. If any random executable did this you would call it malware. No exceptions for a signature.
The historical upload is independently verified from preserved public evidence. Live retest confirms repo upload is currently server disabled for the tested account while trace egress continues. All credentials shown are fake canaries.
Receipts: github.com/arafatkatze/grok-…
Olsen retweeted
『透明ドローン？』
塗装で隠すのではなく、機体全体を高速回転させ、人間の目には背景に溶ける薄いぼかしのように見せる。
news.northwestern.edu/storie…
#drone #UAV #robotics #disappear #RobotDesign #HumanPerception #EnvironmentalMonitoring #PhantomTwist #NorthwesternU
Olsen retweeted
Announcing Grok 4.5, our first model trained specifically for coding and agents. It was trained with Cursor and offers frontier intelligence at leading speeds and cost efficiency.
x.ai/news/grok-4-5
Olsen retweeted
~20 engineers are currently powering X ads (modeling + product + infra). We’re a startup operating in billion+ scale. If you’re looking for outsized impact, end to end full stack work (discover, ship, and sell), a talent dense team, and lots of fun you should join us.
We're hiring exceptional engineers to integrate Grok across our Ads Platform! If you’ve built massive-scale advertising systems in bidding, auctions, marketplace, ranking, prediction & product, let’s transform ads with xAI.
👇Click to apply!
job-boards.greenhouse.io/xai…
If you are not advertising on X, you are shooting yourself in the foot.
I have been working in marketing and advertising since 2016
My honest take:
Twitter (then) wasn't favorable as a platform to advertise most services and products. Especially products. And it was costly to run a campaign - to what end anyway? Wasn't worth it.
X is now seeing a more diverse user base. They also spend more time on the app (all data proven) + the cost to run a campaign has dropped significantly.
I believe X is a viable channel for running campaigns to advertise most services and products.
The key now is letting the advertisers find their right audience
Surprisingly the hardest problems in software are not the complicated ones. The hardest problems are those that are convoluted asf
Complicated problems are very interesting, but convoluted problems suck your soul right out of you
There is a very pure sense of accomplishment in working day and night to solve a complicated problem in an elegant way. So pushing through is fairly easy and enjoyable.
On the other end, having to decipher dumb heuristic decision trees like:
“if this boolean is an artifact of the migration from 2014, ignore it but then it was created by someone whose name starts with the letter X then respect it. But then, if there was a full moon on the day it was created then roll a dice” because of some dumb decision someone made to get shit off their plate a decade ago. Would make the most hardcore eng say “kms”
Worst part of convoluted problems is discovering requirements as you build through very loud feedback.